hns_roce_mr.c

   1/*
   2 * Copyright (c) 2016 Hisilicon Limited.
   3 * Copyright (c) 2007, 2008 Mellanox Technologies. All rights reserved.
   4 *
   5 * This software is available to you under a choice of one of two
   6 * licenses.  You may choose to be licensed under the terms of the GNU
   7 * General Public License (GPL) Version 2, available from the file
   8 * COPYING in the main directory of this source tree, or the
   9 * OpenIB.org BSD license below:
  10 *
  11 *     Redistribution and use in source and binary forms, with or
  12 *     without modification, are permitted provided that the following
  13 *     conditions are met:
  14 *
  15 *      - Redistributions of source code must retain the above
  16 *        copyright notice, this list of conditions and the following
  17 *        disclaimer.
  18 *
  19 *      - Redistributions in binary form must reproduce the above
  20 *        copyright notice, this list of conditions and the following
  21 *        disclaimer in the documentation and/or other materials
  22 *        provided with the distribution.
  23 *
  24 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
  25 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
  26 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
  27 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
  28 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
  29 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
  30 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  31 * SOFTWARE.
  32 */
  33
  34#include <linux/vmalloc.h>
  35#include <linux/count_zeros.h>
  36#include <rdma/ib_umem.h>
  37#include <linux/math.h>
  38#include "hns_roce_device.h"
  39#include "hns_roce_cmd.h"
  40#include "hns_roce_hem.h"
  41#include "hns_roce_trace.h"
  42
  43static u32 hw_index_to_key(int ind)
  44{
  45	return ((u32)ind >> 24) | ((u32)ind << 8);
  46}
  47
  48unsigned long key_to_hw_index(u32 key)
  49{
  50	return (key << 24) | (key >> 8);
  51}
  52
  53static int alloc_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
  54{
  55	struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
  56	struct ib_device *ibdev = &hr_dev->ib_dev;
  57	int err;
  58	int id;
  59
  60	/* Allocate a key for mr from mr_table */
  61	id = ida_alloc_range(&mtpt_ida->ida, mtpt_ida->min, mtpt_ida->max,
  62			     GFP_KERNEL);
  63	if (id < 0) {
  64		ibdev_err(ibdev, "failed to alloc id for MR key, id(%d)\n", id);
  65		return -ENOMEM;
  66	}
  67
  68	mr->key = hw_index_to_key(id); /* MR key */
  69
  70	err = hns_roce_table_get(hr_dev, &hr_dev->mr_table.mtpt_table,
  71				 (unsigned long)id);
  72	if (err) {
  73		ibdev_err(ibdev, "failed to alloc mtpt, ret = %d.\n", err);
  74		goto err_free_bitmap;
  75	}
  76
  77	return 0;
  78err_free_bitmap:
  79	ida_free(&mtpt_ida->ida, id);
  80	return err;
  81}
  82
  83static void free_mr_key(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
  84{
  85	unsigned long obj = key_to_hw_index(mr->key);
  86
  87	hns_roce_table_put(hr_dev, &hr_dev->mr_table.mtpt_table, obj);
  88	ida_free(&hr_dev->mr_table.mtpt_ida.ida, (int)obj);
  89}
  90
  91static int alloc_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr,
  92			struct ib_udata *udata, u64 start)
  93{
  94	struct ib_device *ibdev = &hr_dev->ib_dev;
  95	bool is_fast = mr->type == MR_TYPE_FRMR;
  96	struct hns_roce_buf_attr buf_attr = {};
  97	int err;
  98
  99	mr->pbl_hop_num = is_fast ? 1 : hr_dev->caps.pbl_hop_num;
 100	buf_attr.page_shift = is_fast ? PAGE_SHIFT :
 101			      hr_dev->caps.pbl_buf_pg_sz + PAGE_SHIFT;
 102	buf_attr.region[0].size = mr->size;
 103	buf_attr.region[0].hopnum = mr->pbl_hop_num;
 104	buf_attr.region_count = 1;
 105	buf_attr.user_access = mr->access;
 106	/* fast MR's buffer is alloced before mapping, not at creation */
 107	buf_attr.mtt_only = is_fast;
 108	buf_attr.iova = mr->iova;
 109	/* pagesize and hopnum is fixed for fast MR */
 110	buf_attr.adaptive = !is_fast;
 111	buf_attr.type = MTR_PBL;
 112
 113	err = hns_roce_mtr_create(hr_dev, &mr->pbl_mtr, &buf_attr,
 114				  hr_dev->caps.pbl_ba_pg_sz + PAGE_SHIFT,
 115				  udata, start);
 116	if (err) {
 117		ibdev_err(ibdev, "failed to alloc pbl mtr, ret = %d.\n", err);
 118		return err;
 119	}
 120
 121	mr->npages = mr->pbl_mtr.hem_cfg.buf_pg_count;
 122	mr->pbl_hop_num = buf_attr.region[0].hopnum;
 123
 124	return err;
 125}
 126
 127static void free_mr_pbl(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
 128{
 129	hns_roce_mtr_destroy(hr_dev, &mr->pbl_mtr);
 130}
 131
 132static void hns_roce_mr_free(struct hns_roce_dev *hr_dev, struct hns_roce_mr *mr)
 133{
 134	struct ib_device *ibdev = &hr_dev->ib_dev;
 135	int ret;
 136
 137	if (mr->enabled) {
 138		ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
 139					      key_to_hw_index(mr->key) &
 140					      (hr_dev->caps.num_mtpts - 1));
 141		if (ret)
 142			ibdev_warn_ratelimited(ibdev, "failed to destroy mpt, ret = %d.\n",
 143					       ret);
 144	}
 145
 146	free_mr_pbl(hr_dev, mr);
 147	free_mr_key(hr_dev, mr);
 148}
 149
 150static int hns_roce_mr_enable(struct hns_roce_dev *hr_dev,
 151			      struct hns_roce_mr *mr)
 152{
 153	unsigned long mtpt_idx = key_to_hw_index(mr->key);
 154	struct hns_roce_cmd_mailbox *mailbox;
 155	struct device *dev = hr_dev->dev;
 156	int ret;
 157
 158	/* Allocate mailbox memory */
 159	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
 160	if (IS_ERR(mailbox))
 161		return PTR_ERR(mailbox);
 162
 163	trace_hns_mr(mr);
 164	if (mr->type != MR_TYPE_FRMR)
 165		ret = hr_dev->hw->write_mtpt(hr_dev, mailbox->buf, mr);
 166	else
 167		ret = hr_dev->hw->frmr_write_mtpt(mailbox->buf, mr);
 168	if (ret) {
 169		dev_err(dev, "failed to write mtpt, ret = %d.\n", ret);
 170		goto err_page;
 171	}
 172
 173	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
 174				     mtpt_idx & (hr_dev->caps.num_mtpts - 1));
 175	if (ret) {
 176		dev_err(dev, "failed to create mpt, ret = %d.\n", ret);
 177		goto err_page;
 178	}
 179
 180	mr->enabled = 1;
 181
 182err_page:
 183	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
 184
 185	return ret;
 186}
 187
 188void hns_roce_init_mr_table(struct hns_roce_dev *hr_dev)
 189{
 190	struct hns_roce_ida *mtpt_ida = &hr_dev->mr_table.mtpt_ida;
 191
 192	ida_init(&mtpt_ida->ida);
 193	mtpt_ida->max = hr_dev->caps.num_mtpts - 1;
 194	mtpt_ida->min = hr_dev->caps.reserved_mrws;
 195}
 196
 197struct ib_mr *hns_roce_get_dma_mr(struct ib_pd *pd, int acc)
 198{
 199	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
 200	struct hns_roce_mr *mr;
 201	int ret;
 202
 203	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 204	if (!mr)
 205		return  ERR_PTR(-ENOMEM);
 206
 207	mr->type = MR_TYPE_DMA;
 208	mr->pd = to_hr_pd(pd)->pdn;
 209	mr->access = acc;
 210
 211	/* Allocate memory region key */
 212	hns_roce_hem_list_init(&mr->pbl_mtr.hem_list);
 213	ret = alloc_mr_key(hr_dev, mr);
 214	if (ret)
 215		goto err_free;
 216
 217	ret = hns_roce_mr_enable(hr_dev, mr);
 218	if (ret)
 219		goto err_mr;
 220
 221	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
 222
 223	return &mr->ibmr;
 224err_mr:
 225	free_mr_key(hr_dev, mr);
 226
 227err_free:
 228	kfree(mr);
 229	return ERR_PTR(ret);
 230}
 231
 232struct ib_mr *hns_roce_reg_user_mr(struct ib_pd *pd, u64 start, u64 length,
 233				   u64 virt_addr, int access_flags,
 234				   struct ib_dmah *dmah,
 235				   struct ib_udata *udata)
 236{
 237	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
 238	struct hns_roce_mr *mr;
 239	int ret;
 240
 241	if (dmah) {
 242		ret = -EOPNOTSUPP;
 243		goto err_out;
 244	}
 245
 246	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 247	if (!mr) {
 248		ret = -ENOMEM;
 249		goto err_out;
 250	}
 251
 252	mr->iova = virt_addr;
 253	mr->size = length;
 254	mr->pd = to_hr_pd(pd)->pdn;
 255	mr->access = access_flags;
 256	mr->type = MR_TYPE_MR;
 257
 258	ret = alloc_mr_key(hr_dev, mr);
 259	if (ret)
 260		goto err_alloc_mr;
 261
 262	ret = alloc_mr_pbl(hr_dev, mr, udata, start);
 263	if (ret)
 264		goto err_alloc_key;
 265
 266	ret = hns_roce_mr_enable(hr_dev, mr);
 267	if (ret)
 268		goto err_alloc_pbl;
 269
 270	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
 271
 272	return &mr->ibmr;
 273
 274err_alloc_pbl:
 275	free_mr_pbl(hr_dev, mr);
 276err_alloc_key:
 277	free_mr_key(hr_dev, mr);
 278err_alloc_mr:
 279	kfree(mr);
 280err_out:
 281	atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MR_REG_ERR_CNT]);
 282
 283	return ERR_PTR(ret);
 284}
 285
 286struct ib_mr *hns_roce_rereg_user_mr(struct ib_mr *ibmr, int flags, u64 start,
 287				     u64 length, u64 virt_addr,
 288				     int mr_access_flags, struct ib_pd *pd,
 289				     struct ib_udata *udata)
 290{
 291	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
 292	struct ib_device *ib_dev = &hr_dev->ib_dev;
 293	struct hns_roce_mr *mr = to_hr_mr(ibmr);
 294	struct hns_roce_cmd_mailbox *mailbox;
 295	unsigned long mtpt_idx;
 296	int ret;
 297
 298	if (!mr->enabled) {
 299		ret = -EINVAL;
 300		goto err_out;
 301	}
 302
 303	mailbox = hns_roce_alloc_cmd_mailbox(hr_dev);
 304	ret = PTR_ERR_OR_ZERO(mailbox);
 305	if (ret)
 306		goto err_out;
 307
 308	mtpt_idx = key_to_hw_index(mr->key) & (hr_dev->caps.num_mtpts - 1);
 309
 310	ret = hns_roce_cmd_mbox(hr_dev, 0, mailbox->dma, HNS_ROCE_CMD_QUERY_MPT,
 311				mtpt_idx);
 312	if (ret)
 313		goto free_cmd_mbox;
 314
 315	ret = hns_roce_destroy_hw_ctx(hr_dev, HNS_ROCE_CMD_DESTROY_MPT,
 316				      mtpt_idx);
 317	if (ret)
 318		ibdev_warn(ib_dev, "failed to destroy MPT, ret = %d.\n", ret);
 319
 320	mr->enabled = 0;
 321	mr->iova = virt_addr;
 322	mr->size = length;
 323
 324	if (flags & IB_MR_REREG_PD)
 325		mr->pd = to_hr_pd(pd)->pdn;
 326
 327	if (flags & IB_MR_REREG_ACCESS)
 328		mr->access = mr_access_flags;
 329
 330	if (flags & IB_MR_REREG_TRANS) {
 331		free_mr_pbl(hr_dev, mr);
 332		ret = alloc_mr_pbl(hr_dev, mr, udata, start);
 333		if (ret) {
 334			ibdev_err(ib_dev, "failed to alloc mr PBL, ret = %d.\n",
 335				  ret);
 336			goto free_cmd_mbox;
 337		}
 338	}
 339
 340	ret = hr_dev->hw->rereg_write_mtpt(hr_dev, mr, flags, mailbox->buf);
 341	if (ret) {
 342		ibdev_err(ib_dev, "failed to write mtpt, ret = %d.\n", ret);
 343		goto free_cmd_mbox;
 344	}
 345
 346	ret = hns_roce_create_hw_ctx(hr_dev, mailbox, HNS_ROCE_CMD_CREATE_MPT,
 347				     mtpt_idx);
 348	if (ret) {
 349		ibdev_err(ib_dev, "failed to create MPT, ret = %d.\n", ret);
 350		goto free_cmd_mbox;
 351	}
 352
 353	mr->enabled = 1;
 354
 355free_cmd_mbox:
 356	hns_roce_free_cmd_mailbox(hr_dev, mailbox);
 357
 358err_out:
 359	if (ret) {
 360		atomic64_inc(&hr_dev->dfx_cnt[HNS_ROCE_DFX_MR_REREG_ERR_CNT]);
 361		return ERR_PTR(ret);
 362	}
 363
 364	return NULL;
 365}
 366
 367int hns_roce_dereg_mr(struct ib_mr *ibmr, struct ib_udata *udata)
 368{
 369	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
 370	struct hns_roce_mr *mr = to_hr_mr(ibmr);
 371
 372	if (hr_dev->hw->dereg_mr)
 373		hr_dev->hw->dereg_mr(hr_dev);
 374
 375	hns_roce_mr_free(hr_dev, mr);
 376	kfree(mr);
 377
 378	return 0;
 379}
 380
 381struct ib_mr *hns_roce_alloc_mr(struct ib_pd *pd, enum ib_mr_type mr_type,
 382				u32 max_num_sg)
 383{
 384	struct hns_roce_dev *hr_dev = to_hr_dev(pd->device);
 385	struct device *dev = hr_dev->dev;
 386	struct hns_roce_mr *mr;
 387	int ret;
 388
 389	if (mr_type != IB_MR_TYPE_MEM_REG)
 390		return ERR_PTR(-EINVAL);
 391
 392	if (max_num_sg > HNS_ROCE_FRMR_MAX_PA) {
 393		dev_err(dev, "max_num_sg larger than %d\n",
 394			HNS_ROCE_FRMR_MAX_PA);
 395		return ERR_PTR(-EINVAL);
 396	}
 397
 398	mr = kzalloc(sizeof(*mr), GFP_KERNEL);
 399	if (!mr)
 400		return ERR_PTR(-ENOMEM);
 401
 402	mr->type = MR_TYPE_FRMR;
 403	mr->pd = to_hr_pd(pd)->pdn;
 404	mr->size = max_num_sg * (1 << PAGE_SHIFT);
 405
 406	/* Allocate memory region key */
 407	ret = alloc_mr_key(hr_dev, mr);
 408	if (ret)
 409		goto err_free;
 410
 411	ret = alloc_mr_pbl(hr_dev, mr, NULL, 0);
 412	if (ret)
 413		goto err_key;
 414
 415	ret = hns_roce_mr_enable(hr_dev, mr);
 416	if (ret)
 417		goto err_pbl;
 418
 419	mr->ibmr.rkey = mr->ibmr.lkey = mr->key;
 420	mr->ibmr.length = mr->size;
 421
 422	return &mr->ibmr;
 423
 424err_pbl:
 425	free_mr_pbl(hr_dev, mr);
 426err_key:
 427	free_mr_key(hr_dev, mr);
 428err_free:
 429	kfree(mr);
 430	return ERR_PTR(ret);
 431}
 432
 433static int hns_roce_set_page(struct ib_mr *ibmr, u64 addr)
 434{
 435	struct hns_roce_mr *mr = to_hr_mr(ibmr);
 436
 437	if (likely(mr->npages < mr->pbl_mtr.hem_cfg.buf_pg_count)) {
 438		mr->page_list[mr->npages++] = addr;
 439		return 0;
 440	}
 441
 442	return -ENOBUFS;
 443}
 444
 445int hns_roce_map_mr_sg(struct ib_mr *ibmr, struct scatterlist *sg, int sg_nents,
 446		       unsigned int *sg_offset_p)
 447{
 448	unsigned int sg_offset = sg_offset_p ? *sg_offset_p : 0;
 449	struct hns_roce_dev *hr_dev = to_hr_dev(ibmr->device);
 450	struct ib_device *ibdev = &hr_dev->ib_dev;
 451	struct hns_roce_mr *mr = to_hr_mr(ibmr);
 452	struct hns_roce_mtr *mtr = &mr->pbl_mtr;
 453	int ret, sg_num = 0;
 454
 455	if (!IS_ALIGNED(sg_offset, HNS_ROCE_FRMR_ALIGN_SIZE) ||
 456	    ibmr->page_size < HNS_HW_PAGE_SIZE ||
 457	    ibmr->page_size > HNS_HW_MAX_PAGE_SIZE)
 458		return sg_num;
 459
 460	mr->npages = 0;
 461	mr->page_list = kvcalloc(mr->pbl_mtr.hem_cfg.buf_pg_count,
 462				 sizeof(dma_addr_t), GFP_KERNEL);
 463	if (!mr->page_list)
 464		return sg_num;
 465
 466	sg_num = ib_sg_to_pages(ibmr, sg, sg_nents, sg_offset_p, hns_roce_set_page);
 467	if (sg_num < 1) {
 468		ibdev_err(ibdev, "failed to store sg pages %u %u, cnt = %d.\n",
 469			  mr->npages, mr->pbl_mtr.hem_cfg.buf_pg_count, sg_num);
 470		goto err_page_list;
 471	}
 472
 473	mtr->hem_cfg.region[0].offset = 0;
 474	mtr->hem_cfg.region[0].count = mr->npages;
 475	mtr->hem_cfg.region[0].hopnum = mr->pbl_hop_num;
 476	mtr->hem_cfg.region_count = 1;
 477	ret = hns_roce_mtr_map(hr_dev, mtr, mr->page_list, mr->npages);
 478	if (ret) {
 479		ibdev_err(ibdev, "failed to map sg mtr, ret = %d.\n", ret);
 480		sg_num = 0;
 481	} else {
 482		mr->pbl_mtr.hem_cfg.buf_pg_shift = (u32)ilog2(ibmr->page_size);
 483	}
 484
 485err_page_list:
 486	kvfree(mr->page_list);
 487	mr->page_list = NULL;
 488
 489	return sg_num;
 490}
 491
 492static int mtr_map_region(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 493			  struct hns_roce_buf_region *region, dma_addr_t *pages,
 494			  int max_count)
 495{
 496	int count, npage;
 497	int offset, end;
 498	__le64 *mtts;
 499	u64 addr;
 500	int i;
 501
 502	offset = region->offset;
 503	end = offset + region->count;
 504	npage = 0;
 505	while (offset < end && npage < max_count) {
 506		count = 0;
 507		mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
 508						  offset, &count);
 509		if (!mtts)
 510			return -ENOBUFS;
 511
 512		for (i = 0; i < count && npage < max_count; i++) {
 513			addr = pages[npage];
 514
 515			mtts[i] = cpu_to_le64(addr);
 516			npage++;
 517		}
 518		offset += count;
 519	}
 520
 521	return npage;
 522}
 523
 524static inline bool mtr_has_mtt(struct hns_roce_buf_attr *attr)
 525{
 526	int i;
 527
 528	for (i = 0; i < attr->region_count; i++)
 529		if (attr->region[i].hopnum != HNS_ROCE_HOP_NUM_0 &&
 530		    attr->region[i].hopnum > 0)
 531			return true;
 532
 533	/* because the mtr only one root base address, when hopnum is 0 means
 534	 * root base address equals the first buffer address, thus all alloced
 535	 * memory must in a continuous space accessed by direct mode.
 536	 */
 537	return false;
 538}
 539
 540static inline size_t mtr_bufs_size(struct hns_roce_buf_attr *attr)
 541{
 542	size_t size = 0;
 543	int i;
 544
 545	for (i = 0; i < attr->region_count; i++)
 546		size += attr->region[i].size;
 547
 548	return size;
 549}
 550
 551/*
 552 * check the given pages in continuous address space
 553 * Returns 0 on success, or the error page num.
 554 */
 555static inline int mtr_check_direct_pages(dma_addr_t *pages, int page_count,
 556					 unsigned int page_shift)
 557{
 558	size_t page_size = 1 << page_shift;
 559	int i;
 560
 561	for (i = 1; i < page_count; i++)
 562		if (pages[i] - pages[i - 1] != page_size)
 563			return i;
 564
 565	return 0;
 566}
 567
 568static void mtr_free_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
 569{
 570	/* release user buffers */
 571	if (mtr->umem) {
 572		ib_umem_release(mtr->umem);
 573		mtr->umem = NULL;
 574	}
 575
 576	/* release kernel buffers */
 577	if (mtr->kmem) {
 578		hns_roce_buf_free(hr_dev, mtr->kmem);
 579		mtr->kmem = NULL;
 580	}
 581}
 582
 583static int mtr_alloc_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 584			  struct hns_roce_buf_attr *buf_attr,
 585			  struct ib_udata *udata, unsigned long user_addr)
 586{
 587	struct ib_device *ibdev = &hr_dev->ib_dev;
 588	size_t total_size;
 589
 590	total_size = mtr_bufs_size(buf_attr);
 591
 592	if (udata) {
 593		mtr->kmem = NULL;
 594		mtr->umem = ib_umem_get(ibdev, user_addr, total_size,
 595					buf_attr->user_access);
 596		if (IS_ERR(mtr->umem)) {
 597			ibdev_err(ibdev, "failed to get umem, ret = %pe.\n",
 598				  mtr->umem);
 599			return -ENOMEM;
 600		}
 601	} else {
 602		mtr->umem = NULL;
 603		mtr->kmem = hns_roce_buf_alloc(hr_dev, total_size,
 604					       buf_attr->page_shift,
 605					       !mtr_has_mtt(buf_attr) ?
 606					       HNS_ROCE_BUF_DIRECT : 0);
 607		if (IS_ERR(mtr->kmem)) {
 608			ibdev_err(ibdev, "failed to alloc kmem, ret = %pe.\n",
 609				  mtr->kmem);
 610			return PTR_ERR(mtr->kmem);
 611		}
 612	}
 613
 614	return 0;
 615}
 616
 617static int cal_mtr_pg_cnt(struct hns_roce_mtr *mtr)
 618{
 619	struct hns_roce_buf_region *region;
 620	int page_cnt = 0;
 621	int i;
 622
 623	for (i = 0; i < mtr->hem_cfg.region_count; i++) {
 624		region = &mtr->hem_cfg.region[i];
 625		page_cnt += region->count;
 626	}
 627
 628	return page_cnt;
 629}
 630
 631static bool need_split_huge_page(struct hns_roce_mtr *mtr)
 632{
 633	/* When HEM buffer uses 0-level addressing, the page size is
 634	 * equal to the whole buffer size. If the current MTR has multiple
 635	 * regions, we split the buffer into small pages(4k, required by hns
 636	 * ROCEE). These pages will be used in multiple regions.
 637	 */
 638	return mtr->hem_cfg.is_direct && mtr->hem_cfg.region_count > 1;
 639}
 640
 641static int mtr_map_bufs(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
 642{
 643	struct ib_device *ibdev = &hr_dev->ib_dev;
 644	int page_count = cal_mtr_pg_cnt(mtr);
 645	unsigned int page_shift;
 646	dma_addr_t *pages;
 647	int npage;
 648	int ret;
 649
 650	page_shift = need_split_huge_page(mtr) ? HNS_HW_PAGE_SHIFT :
 651						 mtr->hem_cfg.buf_pg_shift;
 652	/* alloc a tmp array to store buffer's dma address */
 653	pages = kvcalloc(page_count, sizeof(dma_addr_t), GFP_KERNEL);
 654	if (!pages)
 655		return -ENOMEM;
 656
 657	if (mtr->umem)
 658		npage = hns_roce_get_umem_bufs(pages, page_count,
 659					       mtr->umem, page_shift);
 660	else
 661		npage = hns_roce_get_kmem_bufs(hr_dev, pages, page_count,
 662					       mtr->kmem, page_shift);
 663
 664	if (npage != page_count) {
 665		ibdev_err(ibdev, "failed to get mtr page %d != %d.\n", npage,
 666			  page_count);
 667		ret = -ENOBUFS;
 668		goto err_alloc_list;
 669	}
 670
 671	if (need_split_huge_page(mtr) && npage > 1) {
 672		ret = mtr_check_direct_pages(pages, npage, page_shift);
 673		if (ret) {
 674			ibdev_err(ibdev, "failed to check %s page: %d / %d.\n",
 675				  mtr->umem ? "umtr" : "kmtr", ret, npage);
 676			ret = -ENOBUFS;
 677			goto err_alloc_list;
 678		}
 679	}
 680
 681	ret = hns_roce_mtr_map(hr_dev, mtr, pages, page_count);
 682	if (ret)
 683		ibdev_err(ibdev, "failed to map mtr pages, ret = %d.\n", ret);
 684
 685err_alloc_list:
 686	kvfree(pages);
 687
 688	return ret;
 689}
 690
 691int hns_roce_mtr_map(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 692		     dma_addr_t *pages, unsigned int page_cnt)
 693{
 694	struct ib_device *ibdev = &hr_dev->ib_dev;
 695	struct hns_roce_buf_region *r;
 696	unsigned int i, mapped_cnt;
 697	int ret = 0;
 698
 699	/*
 700	 * Only use the first page address as root ba when hopnum is 0, this
 701	 * is because the addresses of all pages are consecutive in this case.
 702	 */
 703	if (mtr->hem_cfg.is_direct) {
 704		mtr->hem_cfg.root_ba = pages[0];
 705		return 0;
 706	}
 707
 708	for (i = 0, mapped_cnt = 0; i < mtr->hem_cfg.region_count &&
 709	     mapped_cnt < page_cnt; i++) {
 710		r = &mtr->hem_cfg.region[i];
 711
 712		if (r->offset + r->count > page_cnt) {
 713			ret = -EINVAL;
 714			ibdev_err(ibdev,
 715				  "failed to check mtr%u count %u + %u > %u.\n",
 716				  i, r->offset, r->count, page_cnt);
 717			return ret;
 718		}
 719
 720		ret = mtr_map_region(hr_dev, mtr, r, &pages[r->offset],
 721				     page_cnt - mapped_cnt);
 722		if (ret < 0) {
 723			ibdev_err(ibdev,
 724				  "failed to map mtr%u offset %u, ret = %d.\n",
 725				  i, r->offset, ret);
 726			return ret;
 727		}
 728		mapped_cnt += ret;
 729		ret = 0;
 730	}
 731
 732	if (mapped_cnt < page_cnt) {
 733		ret = -ENOBUFS;
 734		ibdev_err(ibdev, "failed to map mtr pages count: %u < %u.\n",
 735			  mapped_cnt, page_cnt);
 736	}
 737
 738	return ret;
 739}
 740
 741static int hns_roce_get_direct_addr_mtt(struct hns_roce_hem_cfg *cfg,
 742					u32 start_index, u64 *mtt_buf,
 743					int mtt_cnt)
 744{
 745	int mtt_count;
 746	int total = 0;
 747	u32 npage;
 748	u64 addr;
 749
 750	if (mtt_cnt > cfg->region_count)
 751		return -EINVAL;
 752
 753	for (mtt_count = 0; mtt_count < cfg->region_count && total < mtt_cnt;
 754	     mtt_count++) {
 755		npage = cfg->region[mtt_count].offset;
 756		if (npage < start_index)
 757			continue;
 758
 759		addr = cfg->root_ba + (npage << HNS_HW_PAGE_SHIFT);
 760		mtt_buf[total] = addr;
 761
 762		total++;
 763	}
 764
 765	if (!total)
 766		return -ENOENT;
 767
 768	return 0;
 769}
 770
 771static int hns_roce_get_mhop_mtt(struct hns_roce_dev *hr_dev,
 772				 struct hns_roce_mtr *mtr, u32 start_index,
 773				 u64 *mtt_buf, int mtt_cnt)
 774{
 775	int left = mtt_cnt;
 776	int total = 0;
 777	int mtt_count;
 778	__le64 *mtts;
 779	u32 npage;
 780
 781	while (left > 0) {
 782		mtt_count = 0;
 783		mtts = hns_roce_hem_list_find_mtt(hr_dev, &mtr->hem_list,
 784						  start_index + total,
 785						  &mtt_count);
 786		if (!mtts || !mtt_count)
 787			break;
 788
 789		npage = min(mtt_count, left);
 790		left -= npage;
 791		for (mtt_count = 0; mtt_count < npage; mtt_count++)
 792			mtt_buf[total++] = le64_to_cpu(mtts[mtt_count]);
 793	}
 794
 795	if (!total)
 796		return -ENOENT;
 797
 798	return 0;
 799}
 800
 801int hns_roce_mtr_find(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 802		      u32 offset, u64 *mtt_buf, int mtt_max)
 803{
 804	struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
 805	u32 start_index;
 806	int ret;
 807
 808	if (!mtt_buf || mtt_max < 1)
 809		return -EINVAL;
 810
 811	/* no mtt memory in direct mode, so just return the buffer address */
 812	if (cfg->is_direct) {
 813		start_index = offset >> HNS_HW_PAGE_SHIFT;
 814		ret = hns_roce_get_direct_addr_mtt(cfg, start_index,
 815						   mtt_buf, mtt_max);
 816	} else {
 817		start_index = offset >> cfg->buf_pg_shift;
 818		ret = hns_roce_get_mhop_mtt(hr_dev, mtr, start_index,
 819					    mtt_buf, mtt_max);
 820	}
 821	return ret;
 822}
 823
 824static int get_best_page_shift(struct hns_roce_dev *hr_dev,
 825			       struct hns_roce_mtr *mtr,
 826			       struct hns_roce_buf_attr *buf_attr)
 827{
 828	unsigned int page_sz;
 829
 830	if (!buf_attr->adaptive || buf_attr->type != MTR_PBL || !mtr->umem)
 831		return 0;
 832
 833	page_sz = ib_umem_find_best_pgsz(mtr->umem,
 834					 hr_dev->caps.page_size_cap,
 835					 buf_attr->iova);
 836	if (!page_sz)
 837		return -EINVAL;
 838
 839	buf_attr->page_shift = order_base_2(page_sz);
 840	return 0;
 841}
 842
 843static int get_best_hop_num(struct hns_roce_dev *hr_dev,
 844			    struct hns_roce_mtr *mtr,
 845			    struct hns_roce_buf_attr *buf_attr,
 846			    unsigned int ba_pg_shift)
 847{
 848#define INVALID_HOPNUM -1
 849#define MIN_BA_CNT 1
 850	size_t buf_pg_sz = 1 << buf_attr->page_shift;
 851	struct ib_device *ibdev = &hr_dev->ib_dev;
 852	size_t ba_pg_sz = 1 << ba_pg_shift;
 853	int hop_num = INVALID_HOPNUM;
 854	size_t unit = MIN_BA_CNT;
 855	size_t ba_cnt;
 856	int j;
 857
 858	if (!buf_attr->adaptive || buf_attr->type != MTR_PBL)
 859		return 0;
 860
 861	/* Caculating the number of buf pages, each buf page need a BA */
 862	if (mtr->umem)
 863		ba_cnt = ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz);
 864	else
 865		ba_cnt = DIV_ROUND_UP(buf_attr->region[0].size, buf_pg_sz);
 866
 867	for (j = 0; j <= HNS_ROCE_MAX_HOP_NUM; j++) {
 868		if (ba_cnt <= unit) {
 869			hop_num = j;
 870			break;
 871		}
 872		/* Number of BAs can be represented at per hop */
 873		unit *= ba_pg_sz / BA_BYTE_LEN;
 874	}
 875
 876	if (hop_num < 0) {
 877		ibdev_err(ibdev,
 878			  "failed to calculate a valid hopnum.\n");
 879		return -EINVAL;
 880	}
 881
 882	buf_attr->region[0].hopnum = hop_num;
 883
 884	return 0;
 885}
 886
 887static bool is_buf_attr_valid(struct hns_roce_dev *hr_dev,
 888			      struct hns_roce_buf_attr *attr)
 889{
 890	struct ib_device *ibdev = &hr_dev->ib_dev;
 891
 892	if (attr->region_count > ARRAY_SIZE(attr->region) ||
 893	    attr->region_count < 1 || attr->page_shift < HNS_HW_PAGE_SHIFT) {
 894		ibdev_err(ibdev,
 895			  "invalid buf attr, region count %u, page shift %u.\n",
 896			  attr->region_count, attr->page_shift);
 897		return false;
 898	}
 899
 900	return true;
 901}
 902
 903static int mtr_init_buf_cfg(struct hns_roce_dev *hr_dev,
 904			    struct hns_roce_mtr *mtr,
 905			    struct hns_roce_buf_attr *attr)
 906{
 907	struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
 908	struct hns_roce_buf_region *r;
 909	size_t buf_pg_sz;
 910	size_t buf_size;
 911	int page_cnt, i;
 912	u64 pgoff = 0;
 913
 914	if (!is_buf_attr_valid(hr_dev, attr))
 915		return -EINVAL;
 916
 917	/* If mtt is disabled, all pages must be within a continuous range */
 918	cfg->is_direct = !mtr_has_mtt(attr);
 919	cfg->region_count = attr->region_count;
 920	buf_size = mtr_bufs_size(attr);
 921	if (need_split_huge_page(mtr)) {
 922		buf_pg_sz = HNS_HW_PAGE_SIZE;
 923		cfg->buf_pg_count = 1;
 924		/* The ROCEE requires the page size to be 4K * 2 ^ N. */
 925		cfg->buf_pg_shift = HNS_HW_PAGE_SHIFT +
 926			order_base_2(DIV_ROUND_UP(buf_size, HNS_HW_PAGE_SIZE));
 927	} else {
 928		buf_pg_sz = 1 << attr->page_shift;
 929		cfg->buf_pg_count = mtr->umem ?
 930			ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz) :
 931			DIV_ROUND_UP(buf_size, buf_pg_sz);
 932		cfg->buf_pg_shift = attr->page_shift;
 933		pgoff = mtr->umem ? mtr->umem->address & ~PAGE_MASK : 0;
 934	}
 935
 936	/* Convert buffer size to page index and page count for each region and
 937	 * the buffer's offset needs to be appended to the first region.
 938	 */
 939	for (page_cnt = 0, i = 0; i < attr->region_count; i++) {
 940		r = &cfg->region[i];
 941		r->offset = page_cnt;
 942		buf_size = hr_hw_page_align(attr->region[i].size + pgoff);
 943		if (attr->type == MTR_PBL && mtr->umem)
 944			r->count = ib_umem_num_dma_blocks(mtr->umem, buf_pg_sz);
 945		else
 946			r->count = DIV_ROUND_UP(buf_size, buf_pg_sz);
 947
 948		pgoff = 0;
 949		page_cnt += r->count;
 950		r->hopnum = to_hr_hem_hopnum(attr->region[i].hopnum, r->count);
 951	}
 952
 953	return 0;
 954}
 955
 956static u64 cal_pages_per_l1ba(unsigned int ba_per_bt, unsigned int hopnum)
 957{
 958	return int_pow(ba_per_bt, hopnum - 1);
 959}
 960
 961static unsigned int cal_best_bt_pg_sz(struct hns_roce_dev *hr_dev,
 962				      struct hns_roce_mtr *mtr,
 963				      unsigned int pg_shift)
 964{
 965	unsigned long cap = hr_dev->caps.page_size_cap;
 966	struct hns_roce_buf_region *re;
 967	unsigned int pgs_per_l1ba;
 968	unsigned int ba_per_bt;
 969	unsigned int ba_num;
 970	int i;
 971
 972	for_each_set_bit_from(pg_shift, &cap, sizeof(cap) * BITS_PER_BYTE) {
 973		if (!(BIT(pg_shift) & cap))
 974			continue;
 975
 976		ba_per_bt = BIT(pg_shift) / BA_BYTE_LEN;
 977		ba_num = 0;
 978		for (i = 0; i < mtr->hem_cfg.region_count; i++) {
 979			re = &mtr->hem_cfg.region[i];
 980			if (re->hopnum == 0)
 981				continue;
 982
 983			pgs_per_l1ba = cal_pages_per_l1ba(ba_per_bt, re->hopnum);
 984			ba_num += DIV_ROUND_UP(re->count, pgs_per_l1ba);
 985		}
 986
 987		if (ba_num <= ba_per_bt)
 988			return pg_shift;
 989	}
 990
 991	return 0;
 992}
 993
 994static int mtr_alloc_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
 995			 unsigned int ba_page_shift)
 996{
 997	struct hns_roce_hem_cfg *cfg = &mtr->hem_cfg;
 998	int ret;
 999
1000	hns_roce_hem_list_init(&mtr->hem_list);
1001	if (!cfg->is_direct) {
1002		ba_page_shift = cal_best_bt_pg_sz(hr_dev, mtr, ba_page_shift);
1003		if (!ba_page_shift)
1004			return -ERANGE;
1005
1006		ret = hns_roce_hem_list_request(hr_dev, &mtr->hem_list,
1007						cfg->region, cfg->region_count,
1008						ba_page_shift);
1009		if (ret)
1010			return ret;
1011		cfg->root_ba = mtr->hem_list.root_ba;
1012		cfg->ba_pg_shift = ba_page_shift;
1013	} else {
1014		cfg->ba_pg_shift = cfg->buf_pg_shift;
1015	}
1016
1017	return 0;
1018}
1019
1020static void mtr_free_mtt(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
1021{
1022	hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
1023}
1024
1025/**
1026 * hns_roce_mtr_create - Create hns memory translate region.
1027 *
1028 * @hr_dev: RoCE device struct pointer
1029 * @mtr: memory translate region
1030 * @buf_attr: buffer attribute for creating mtr
1031 * @ba_page_shift: page shift for multi-hop base address table
1032 * @udata: user space context, if it's NULL, means kernel space
1033 * @user_addr: userspace virtual address to start at
1034 */
1035int hns_roce_mtr_create(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr,
1036			struct hns_roce_buf_attr *buf_attr,
1037			unsigned int ba_page_shift, struct ib_udata *udata,
1038			unsigned long user_addr)
1039{
1040	struct ib_device *ibdev = &hr_dev->ib_dev;
1041	int ret;
1042
1043	trace_hns_buf_attr(buf_attr);
1044	/* The caller has its own buffer list and invokes the hns_roce_mtr_map()
1045	 * to finish the MTT configuration.
1046	 */
1047	if (buf_attr->mtt_only) {
1048		mtr->umem = NULL;
1049		mtr->kmem = NULL;
1050	} else {
1051		ret = mtr_alloc_bufs(hr_dev, mtr, buf_attr, udata, user_addr);
1052		if (ret) {
1053			ibdev_err(ibdev,
1054				  "failed to alloc mtr bufs, ret = %d.\n", ret);
1055			return ret;
1056		}
1057
1058		ret = get_best_page_shift(hr_dev, mtr, buf_attr);
1059		if (ret)
1060			goto err_init_buf;
1061
1062		ret = get_best_hop_num(hr_dev, mtr, buf_attr, ba_page_shift);
1063		if (ret)
1064			goto err_init_buf;
1065	}
1066
1067	ret = mtr_init_buf_cfg(hr_dev, mtr, buf_attr);
1068	if (ret)
1069		goto err_init_buf;
1070
1071	ret = mtr_alloc_mtt(hr_dev, mtr, ba_page_shift);
1072	if (ret) {
1073		ibdev_err(ibdev, "failed to alloc mtr mtt, ret = %d.\n", ret);
1074		goto err_init_buf;
1075	}
1076
1077	if (buf_attr->mtt_only)
1078		return 0;
1079
1080	/* Write buffer's dma address to MTT */
1081	ret = mtr_map_bufs(hr_dev, mtr);
1082	if (ret) {
1083		ibdev_err(ibdev, "failed to map mtr bufs, ret = %d.\n", ret);
1084		goto err_alloc_mtt;
1085	}
1086
1087	return 0;
1088
1089err_alloc_mtt:
1090	mtr_free_mtt(hr_dev, mtr);
1091err_init_buf:
1092	mtr_free_bufs(hr_dev, mtr);
1093
1094	return ret;
1095}
1096
1097void hns_roce_mtr_destroy(struct hns_roce_dev *hr_dev, struct hns_roce_mtr *mtr)
1098{
1099	/* release multi-hop addressing resource */
1100	hns_roce_hem_list_release(hr_dev, &mtr->hem_list);
1101
1102	/* free buffers */
1103	mtr_free_bufs(hr_dev, mtr);
1104}